AlphaFold

The Evoformer is the central neural network module that ingests a Multiple Sequence Alignment (MSA) and a pairwise residue representation. It exchanges information between these two tracks using novel attention mechanisms, allowing the model to reason about evolutionary correlations and spatial proximity simultaneously. This block identifies which residues are likely to be in contact based on residue coevolution patterns, effectively inferring a protein's folded topology before any 3D coordinates are generated.

The Structure Module takes the abstract representations from the Evoformer and produces an explicit 3D structure. It represents each residue as a geometric frame—a rotation and translation—and iteratively refines these frames. A key innovation is Invariant Point Attention (IPA), which operates on 3D point clouds in a way that is SE(3) equivariant: rotating or translating the input coordinates results in an identically transformed output. This ensures the model respects the fundamental symmetries of physical space without needing to learn them from data.

AlphaFold2 employs a recycling mechanism where the model's initial predictions are fed back as input for a subsequent pass. This iterative process allows the network to refine its predictions over multiple cycles, progressively resolving structural details. The outputs from the final Structure Module iteration are used as the refined input for the next full pass through the Evoformer, enabling the model to correct initial errors and converge on a more accurate structure. This is typically performed three times, with most gains occurring in the first recycle.

AlphaFold outputs two critical per-residue and pairwise confidence metrics. The Predicted Local Distance Difference Test (pLDDT) scores each residue from 0 to 100, estimating how well the local structure matches an experimental reference. Residues with pLDDT > 90 are considered high-accuracy. The Predicted Aligned Error (PAE) estimates the expected positional error between any two residues. A low PAE between domains indicates a confident relative orientation, while high PAE suggests flexibility or uncertainty. These metrics are essential for determining which parts of a prediction are trustworthy for downstream applications like drug design.

Unlike previous methods that relied on separate, hand-crafted stages, AlphaFold2 is trained end-to-end. The entire pipeline—from MSA processing and Evoformer reasoning to Structure Module coordinate generation—is a single differentiable model. This allows gradients to flow from the final 3D coordinate loss back through every component, enabling the network to learn optimal internal representations for the singular task of structure prediction. The model was trained on structures from the Protein Data Bank (PDB) and achieved groundbreaking results in the CASP14 blind assessment, reaching a median GDT_TS score competitive with experimental methods.

DeepMind released the AlphaFold2 source code and model parameters under an open-source license, catalyzing an explosion of research. They also partnered with the European Bioinformatics Institute to create the AlphaFold Protein Structure Database, which contains over 200 million predicted structures covering nearly every known protein sequence. This resource has democratized access to high-quality structural models, enabling researchers worldwide to investigate protein function, disease mechanisms, and drug targets without needing experimental structures. The ecosystem now includes derivatives like AlphaFold-Multimer for complex prediction and ColabFold for accessible cloud-based execution.